The best prediction programs for C13 are, imo, from ACDLabs (http://acdlabs.com) and Modgraph's NMRPredict (http://www.modgraph.co.uk/). if you're looking for a one of, then Wolfgang Robien's Csearch webservice (http://homepage.univie.ac.at/wolfgang.robien/csearch_server_info.html) would do probably.

DEfinitively the ACD NMR subroutine within ACD Labs

That depends, ACD, Chemdraw and MNOVA are "just" using lookup tables as far as I know. That works reasonably well as long as you molecule is not exotic. Also they do not care about the conformation. You can try ab initio methods (Spartan or Gaussian) that can very useful to get a general idea.

I will strongly suggest to move to DFT calculations with software like Gaussian, nowadays such a calculation can be done in a desktop computer. Just, the typical B3LYP/6-31+g(d,p) level of theory in vacuo GIAO computation will furnish a much more accurate chemical shift prediction for small to medium sized molecules. You can take a look to the review by Petrovic et al. covering this topic

http://www.ingentaconnect.com/content/ben/coc/2010/00000014/00000015/art00007

DFT methods have their place but to claim that they are more accurate than other methods is to gloss-over the difficulties and limitations of this approach. I suggest that you look at the following article (DOI 10.1002/mrc.2571) to see a comparitive review. In general, database methods perform well if the structural fragments of the test structure are well-represented in the database. Neural networks perform well for cases where interpolation is required (i.e. poorly represented fragments). DFT can work well but is severely challenged by conformationally flexible structures. Also, DFT is slow compared with database methods. It is exceptionally slow if you have great flexibility in the molecule (unless you have a large Linux cluster at your disposal)

Hi! I was using ChemDraw for predict NMR chemical shifts. I obtain similar chemical shift compared with experimental NMR.

ACD and MNOVA are both good choices.

ACD and MNOVA will perform well in "standard" molecules. Definitely I will not recommend ChemDraw as a prediction software... it is not intended to be used as it. If those easy methods don't provide a result that matches the experimental data (not surprising) then go to DFT methods, that nowadays are easy and accurate. For example using the Gaussian provided gaussView software you can easily setup basic chemical shift calculations.